Case study · See Stella
The company runs on a brain, and the brain has rules about what counts as true.
See Stella is the business we are building ourselves. The interesting part is not the product. It is that the operation underneath it is run by a system that maintains its own institutional memory.
What is in it
- Notes in the brain
Measured 5 September 2026. The brain never freezes this number in its own prose.
600+ - Retrieval lenses
Actions, decisions, data, building, marketing. Not folders.
5 - Hard rules
Quoted verbatim by every agent that reads or writes.
6 - Scheduled loops
Daily briefing, digest, metric pulse, engineering pulse, usage intelligence, weekly steward.
6
Every growing company loses the same thing, and usually accepts it.
A business generates more context than its people can hold. The same decision gets made twice because nobody remembers the first one. A number gets quoted from memory in a meeting six months after it stopped being true. What a system actually does drifts away from what the documentation says it does, quietly, until someone is bitten by it.
Most companies treat that as the cost of moving fast. It is not. It is a memory problem, and memory problems are solvable if you are willing to be strict about what gets written down and what happens to it afterwards.
Store what is true. Reference what changes.
The brain is a git-tracked knowledge base that agents read from and write to, organised around one keystone rule. Durable knowledge lives as notes. Anything that moves on its own, a metric, a deploy status, a model version, is never written down as a fact. It is stored as a pointer to wherever the live answer lives, and a reading agent fetches it fresh.
That single distinction is what stops a knowledge base rotting. Notes do not go stale because the things that go stale were never put in them.
Six rules, quoted rather than paraphrased.
- Never freeze a live number
Dynamic values point at their source. A note saying “638 of 1,276 scans” is a landmine: the moment it moves, the note lies and nobody notices.
- Never hard-delete
Retired notes are archived with a date prefix, never removed.
- Never silently overwrite
A changed fact is superseded, not edited. The new note goes live before the old one retires, so there is never a moment with no answer.
- Every note declares itself
Status, owner, when it was last reviewed, when it next must be.
- Ignore what has been retired
On retrieval, superseded and archived notes are filtered out rather than trusted.
- Raw goes to the record, not the canon
Captured material lands as a dated event. Only distilled, confirmed truth is promoted.
It maintains itself on a schedule
- Daily briefing and digest
Rolls the day up into one readable place.
- Metric pulse
Writes a dated reading of the numbers, as an event rather than a standing fact.
- Engineering pulse
What shipped and what is in progress, read from the repository.
- Usage intelligence
What people actually did in the product, against what was expected.
- The weekly steward
Reads every note, and for each one past its review date returns keep, update or archive.
Why the steward is the part that matters
Any team can write a review checklist. The reason review checklists fail is that nothing fires them. The steward is an agent on a schedule: it reads every tracked note, finds the ones past their review date or with contradictory metadata, and returns a verdict for each. Keep, and it bumps the review forward. Update, and it opens a task addressed to that note’s owner. Archive, and the note leaves the live set.
It also checks that every superseded note points at whatever replaced it, and that nothing still live links to something retired. Then it writes an audit report. The system is only treated as self-maintaining once that has actually happened on its own schedule, not when someone declares it done.
The honesty is engineered, not asserted
The architecture document names its own failure modes, including the ones it has not solved. The linter runs on a schedule rather than as a commit hook, because the sync tool commits without running hooks and a hook would silently miss most writes. The weekly steward only fires while a particular machine is awake, which is written down as a limitation rather than left for someone to discover.
It also corrects itself in public. The current version records that an earlier version described folders that were never built, and says so in the document rather than quietly deleting the claim. Changing how any of it works requires a decision note explaining why, written in the same change as the edit. The map cannot drift from the territory without leaving a paper trail.
The product it runs
Stella measures skin over time. When someone starts using a product it opens an observation window on the things that product should change, waits the weeks the answer genuinely takes, and then reports back: keep using it, step back from it, or not enough evidence either way. It observes rather than diagnoses, publishes how much a repeat scan moves, and will not show an insight until there is enough behind it.
Those product rules and the brain’s rules are the same rules. A company that refuses to state a number it cannot verify builds a product that refuses to claim a result it cannot support.
Who owns it
We do. See Stella was founded in 2023 by Tanuja and Rishabh with a dermatologist co-founder. It is where we work out how an AI-native operation should run before we bring any of it to a client, and it is the reason we can describe this from the inside rather than from a conference talk.
Where it starts
What does your business know that only one person can find?
A Connected Business Audit maps where the operating knowledge actually sits, and what it would take for the business to hold it rather than the people.
Start a conversation